Parsifal descending the gradient is a neural composition by Philipp Ernst for piano and cello that is based on Wagner’s Parsifal. A neural network extracted the melodic “style” of Parsifal and then generated a new composition. The neural network, however, composed the score still in collaboration with the human, that controlled and tuned up the “temperature” of its improvisation (this is what many journalists refer to when they attempt to say “a music score composed by artificial intelligence”). We are happy to present Ernst’s piece for the PhD ceremony in honour of science fiction writer and cybernetician Herbert W. Franke: interestingly, in 1961, in the same year the first operative neural network was disclosed by Frank Rosenblatt, Franke published the novel The Mind Net.

Technical description: Parsifal descending the gradient is a duet for piano and cello created by two recurrent neural networks. The monophonic cello score is generated by the Tensorflow LSTM Magenta Melody Attention using mostly Richard Wagner’s Parsifal Vorspiel as input data. The Parsifal Vorspiel focuses on the most important Leitmotive of Wagner’s Parsifal. The Liebesmahl-, Glaubens– and Gralmotiv. Those motives which work as a base for most of the Parsifal motives where additionally extracted from the whole opera score to be used as input data and their character can be clearly recognized in the cello score.

To find a polyphonic accompaniment to the cello score other compositions were used, as the Parsifal score itself is musically too complex to train a Neural Network for harmonization. After several experiments a mix of Wagner’s Tannhäuser Ouvertüre and sarabands from Händel to Satie were used to create a polyphonic harmonization. Finally the note values of the accompanying score were manipulated to fit the dynamic of the cello piece. No notes were altered in pitch or position, only the chord lengths was adjusted.

Performed by Uriah Tutter (cello) and Vincent Herrmann (piano).
ZKM Vortragssaal, Karlsruhe, 14 February 2018.